LED-Induced fluorescence and image analysis
to detect stink bug damage in cotton bolls
Mustafic et al.
Mustafic et al. Journal of Biological Engineering 2013, 7:5
http://www.jbioleng.org/content/7/1/5
Mustafic et al. Journal of Biological Engineering 2013, 7:5
http://www.jbioleng.org/content/7/1/5
RE SE A RCH
Open Access
LED-Induced fluorescence and image analysis
to detect stink bug damage in cotton bolls
Adnan Mustafic1 , Erin E Roberts1 , Michael D Toews2 and Mark A Haidekker1*
Abstract
Background: Stink bugs represent a major agricultural pest complex attacking more than 200 wild and cultivated
plants, including cotton in the southeastern US. Stink bug feeding on developing cotton bolls will cause boll abortion
or lint staining and thus reduced yield and lint value. Current methods for stink bug detection involve manual
harvesting and cracking open of a sizable number of immature cotton bolls for visual inspection. This process is
cumbersome, time consuming, and requires a moderate level of experience to obtain accurate estimates. To improve
detection of stink bug feeding, we present here a method based on fluorescent imaging and subsequent image
analyses to determine the likelihood of stink bug damage in cotton bolls.
Results: Damage to different structures of cotton bolls including lint and carpal wall can be observed under blue
LED-induced fluorescence. Generally speaking, damaged regions fluoresce green, whereas non-damaged regions
with chlorophyll fluoresce red. However, similar fluorescence emission is also observable on cotton bolls that have not
been fed upon by stink bugs. Criteria based on fluorescent intensity and the size of the fluorescent spot allow to
differentiate between true positives (fluorescent regions associated with stink bug feeding) and false positives
(fluorescent regions due to other causes). We found a detection rates with two combined criteria of 87% for
true-positive marks and of 8% for false-positive marks.
Conclusions: The imaging technique presented herein gives rise to a possible detection apparatus where a cotton
boll is imaged in the field and images processed by software. The unique fluorescent signature left by stink bugs can
be used to determine with high probability if a cotton boll has been punctured by a stink bug. We believe this
technique, when integrated in a suitable device, could be used for more accurate detection in the field and allow for
more optimized application of pest control.
Background
The southern green stink bug (Nezara viridula L.)
(Hemiptera: Pentatomidae) is one example of a stink bug
that feeds on and thereby damages plants. This species
is thought to originate from Ethiopia, but is now found
in many tropical and subtropical regions of the world.
This species is commonly found in the southern United
States where it is considered a major pest attacking agricultural crops. Stink bugs feed on a variety of wild and
cultivated plants and crops including but not limited to
legumes, nuts, and various fruits and vegetables [1-4]. One
of the extensively planted crops in the southeastern U.S.
susceptible to stink bug infestation is cotton (Gossypium
hirsutum L.) [1]. Most stink bugs feed on developing cotton bolls, which occur after anthesis. Both nymphs and
adults are capable of feeding and damaging crops, and
*Correspondence: mhaidekk@uga.edu
1 College of Engineering, University of Georgia, Athens, GA 30602-4435, USA
Full list of author information is available at the end of the article
mature nymphs tend to cause more damage than younger
instars [5].
The feeding mechanism of stink bugs includes the
extension of a piercing mouthpart (proboscis), with which
it penetrates the outer wall to reach the developing seeds.
Pathogenic bacteria may be introduced during feeding,
thereby adding an additional destructive agent resulting in further degradation of bolls, especially in less
mature bolls [6,7]. Cotton bolls damaged by stink bugs
result in reduced fiber length, quality, and uniformity
[8-10]. Another effect of stink bug feeding on cotton
bolls is the lower rate of seed germination [11]. These
types of damage greatly diminish the value of cotton lint
and can cause substantial impediments to a stable cotton supply. The most common approach to reduce stink
bug damage is to wait until damage exceeds an economic threshold and then apply insecticides to mitigate
the population.
© 2013 Mustafic et al.; licensee BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative
Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and
reproduction in any medium, provided the original work is properly cited.
Mustafic et al. Journal of Biological Engineering 2013, 7:5
http://www.jbioleng.org/content/7/1/5
While highly specific insecticides and genetic engineering of the host crop have become preponderant methods
in agricultural pest control, these methods have not been
effective against stink bugs. Therefore, stink bugs are generally managed using broad spectrum insecticides, such as
pyrethroids and organophosphates, that kill the intended
pests but also many non-target beneficial insects. Proper
timing and application of insecticides is critical to reduce
the risk of secondary pest outbreaks [12,13]. To prevent financial losses, insecticides need to be applied only
when the estimated damage exceeds an economic threshold. Moreover, the choice of insecticide chemistry can
have a profound effect on efficacy. Along with choosing
appropriate types of insecticides for application, timing
is important, because organophosphates and pyrethroids
have limited residual activity in the field. In other words,
insecticides need to only be applied at the correct time and
place or the grower risks erosion of financial gains.
It is of essential importance to continue research in this
field because of the economic importance of cotton. The
United States are the world’s largest exporter of cotton
with 11.6 billion bales per year in 2012, which is nearly
three times more than the next closest country, Australia,
with 4.25 billion bales per year [14]. The southeastern
United States, which is the production region with the
most stink bug damage, produces approximately 25% of
the US upland cotton crop [15].
To develop an appropriate plan for pest control in the
field, two requisite and codependent challenges regarding sampling and damage detection exist. The currently
recommended practice of sampling cotton bolls is time
consuming, burdensome, and destructive to the test population of cotton bolls. Recent studies describe various issues with monitoring stink bug populations which
are predisposed to aggregate in fields and can be difficult to detect with the naked eye [16,17]. Even with
optimized sampling plans, cotton boll examination for
potential damage is an invasive procedure which renders the bolls useless because it requires them to be
cracked open. Development of a new method with proven
accuracy and efficiency to detect stink bug feeding is
necessary.
In this article, we propose an image analysis technique
that builds on previous research by Xia et al. [18]: the
presence of different fluorescence emission was reported,
depending on whether the cotton boll was damaged by
stink bugs or not. Stink bug feeding results in a unique
fluorescent signature on the cotton boll wall that can be
imaged and analyzed for discriminating properties. One of
the advantages of this technique is that it is non-invasive
and nondestructive. Here, we report on an image segmentation and analysis process capable of distinguishing
feeding marks from stink bugs versus those caused by
other factors.
Page 2 of 9
Methods
Stink bug rearing
Southern green stink bugs were reared in the laboratory
on a diet of fresh green beans or okra pods following
the methods of Harris and Todd [19]. Briefly, adult stink
bugs captured in the field were brought into the lab and
placed in 37.9 liter glass aquaria lined with paper towels.
Adults oviposited on the paper towels. These were transferred to ventilated petri dishes until the eggs hatched.
Nymphs were maintained in petri dishes or small plastic wide-mouth jars at 25.0°C and 65% relative humidity.
Adults used for infesting bolls were less than two weeks
old and of mixed sex.
Cotton boll development
Cotton plants were started from seed and grown in the
greenhouse to produce bolls for the study. Three cotton
seeds (FM 9063 B2RF) per 11.35 liter plastic pot were
sewn in Metro Mix 300 growing medium. After germination, two of the plants were culled and the remaining
plant was fertilized monthly with Osmocote 14-14-14 and
Micromax fertilizer (the Scotts Co. LLC, Marsville, OH).
Once the plant began flowering, individual white flowers
were marked daily using flagging tape and then allowed
to develop normally for 10 to 14 days. When the bolls
reached 10 to 14 d past anthesis, a single southern green
stink bug adult was caged in a mesh bag on each boll and
subtending leaf for a period of 48 h.
After treatment with the stink bug, bolls were excised
from the plant and shipped with overnight service to the
Athens campus for imaging. At the time the bolls were
cut from the plant they had an external boll diameter
of 2.3 to 2.5 cm. Control bolls were treated exactly as
described, except that no stink bugs were introduced into
the mesh bags. A total of 136 bolls were harvested that
were exposed to a stink bug, and 109 control cotton bolls
were harvested.
Mechanically punctured cotton bolls
Sixteen additional cotton bolls were punctured with a sterile syringe needle (31 Ga, 8mm long, Beckton-Dickinson,
product no. 328418) as a control. These bolls were treated
exactly as described in the previous section with two
differences: (a) no stink bugs were introduced into the
mesh bag. and (b) before harvesting, the bolls were gently cleaned using rubbing alcohol on a cotton swab to
remove any microorganisms that could contaminate the
boll. Puncturing was done manually, and care was taken
that the needle penetrated into the lint tissue. The punctured bolls were harvested immediately. Bolls with needle
punctures were processed in the same manner as the other
bolls. The purpose of these controls were to examine the
difference between purely mechanical puncture damage
and the specific puncture damage caused by stink bugs.
Mustafic et al. Journal of Biological Engineering 2013, 7:5
http://www.jbioleng.org/content/7/1/5
Imaging apparatus
The imaging apparatus, shown in Figure 1, consisted of
a consumer-grade single lens reflex (SLR) digital camera
(Konica Minolta, Dynax Maxxum 7D, Tokyo, Japan) with
a Sigma 50 mm f/2.8 fixed focus lens. An emission longpass filter (490nm, Chroma Technology Corp., Brattleboro, VT) was positioned on a post in front of the camera.
Two LED sources were available. First, a high-intensity
370nm LED area illuminator (Edmund Optics NT59-369,
center wavelength 370nm, FWHM 35nm, driver current
500mA) provided near-UV light similar to the broad-band
UV source used in [18]. Second, a dual deep blue LED
excitation source (Philips/Luxeon LXHL-LR5C with custom LED driver) provided continuous and homogeneous
excitation light at 440nm with a full-width half-maximum
bandwidth of 20nm. Additional excitation bandpass filters D450/40 (Chroma) were placed on the blue LED to
reduce the spectral overlap between excitation and emission light and limit the emission spectral range to 420nm –
460nm. A rotary stage held cotton bolls during imaging. A
non-fluorescent black cardboard behind the sample stage
reduced background fluorescence and scattered light.
Image acquisition and processing
All images were acquired with lens aperture set at f/11
and an exposure time of 15 seconds. All automatic camera functions were disabled, the white balance was fixed to
5500 K, camera sensitivity kept at ISO 400, and the camera
set to acquire raw images in MRW (Minolta Raw) format. Each cotton boll was placed on the rotary stage and
rotated 90° after each image acquisition, resulting in four
images per cotton boll. Raw-format images have a depth of
12 bits per pixel, and all images were converted with UFRaw (http://ufraw.sourceforge.net) to 16-bit TIFF images
with the 12-bit depth fully retained.
Figure 1 Side view of the imaging apparatus. A SLR camera is
mounted on posts in front of the sample, with a dichroic emission
longpass filter placed between sample and lens. Two blue LED
sources (at 45°angle from the lens-sample path) provide
homogeneous excitation light. The cotton boll sample is placed on a
rotary stage to allow images to be taken from all sides. A
non-fluorescent black cardboard backdrop blocks scattered light and
background fluorescence.
Page 3 of 9
All further image processing was performed with Crystal Image [20], a quantitative image analysis software. The
images were first subjected to a center-weighted median
filter, and their resolution reduced to 1508 by 1004 pixels with a 2 × 2 binning operation. The reduced-scale
images showed a resolution of 18 pixels/mm. After the initial steps, two different paths were taken: Mask creation,
and computation of a ratiometric image. For the mask,
the red channel was blurred with a second-order Butterworth lowpass filter with a cutoff of 10 pixels−1 . Background separation was performed with a fixed threshold,
made possible by the black cloth backdrop. The resulting mask contained the image value 1 for all pixels that
correspond to the cotton boll, and 0 for all background
pixels. The ratiometric image was computed by extracting
the red and green channels from the image and dividing
the green channel by the red channel on a pixel-bypixel basis. Subsequent multiplication of this intermediate image with the mask, also on a pixel-by-pixel basis,
yielded an image that contained zero-valued pixels for the
background and the green/red intensity ratio for the cotton boll pixels. The image processing steps are illustrated
in Figure 2.
Candidates for stink bug puncture marks were identified with a local maximum filter with the a priori
settings of a peak-to-peak distance of 15 or more pixels and a minimum intensity difference of 10 to the
average inside a 15-pixel radius and segmented with a
variant of the greedy snake algorithm [20]. This snake
algorithm is a special variant of active contour models. A circle is placed near the fluorescent mark that
is larger than the actual mark, and whose placement
is uncritical. The circle is discretized into 32 vertices,
and the vertices are iteratively displaced to minimize
the total snake energy. The snake energy is defined
as the sum of the vertex distance (stretching energy),
the path first derivative (bending energy) and the negative image gradient (external energy). With this energy
formulation, the snake tends to contract in a regular (i.e., circular fashion) until it locks onto the image
gradient, and the iterative displacement ends near the
gradient maximum of the boundary of the fluorescent
mark.
Fluorescent marks were excluded from analysis when
their area was below 28 pixels (0.09 mm2 ) or their integrated intensity was below 0.04 mm−2 . Furthermore, all
fluorescent marks with an aspect ratio greater than two or
with a convex shape were excluded. The snake was then
converted into a circle, and the average intensity I2 was
measured inside this circle. A smaller concentric circular
region with a radius of 4 pixels (≈ 0.05 mm2 ) was then
selected to compute the central average intensity I1 . A
ratio I1 /I2 was computed for each candidate for stink bug
puncture marks (Figure 3).
Mustafic et al. Journal of Biological Engineering 2013, 7:5
http://www.jbioleng.org/content/7/1/5
Page 4 of 9
Figure 2 Processing steps of cotton boll images. A native camera raw image (A) is separated into the monochromatic R, G, and B channels, and
the B channel is discarded. To obtain the cotton boll mask, the green channel is filtered and reduced in size (B), then further blurred with a
Butterworth lowpass filter (C). The mask is then obtained by thresholding (D). In parallel, a normalized intensity is generated by dividing the green
channel by the red channel on a pixel-by-pixel basis (E). Multiplication with the mask yields the final image of normalized green intensity from the
cotton boll, separated from background (F). The background in (F) is indicated by the gray checkerboard pattern and is excluded from any image
analysis steps.
Statistical analysis
Statistical analysis was performed in R (http://www.rproject.org). The ability of the area metric and the
intensity metric to distinguish true and false positives
was tested individually with the Kruskal-Wallis test and
Dunn’s multiple comparison test. The choice of the
Kruskal-Wallis test over one-way ANOVA was dictated
by the nonparametric distribution of the data. Deviation
of a sample’s median value from a hypothetical value
was tested with the Wilcoxon signed rank sum test. A
significance level of α=0.05 was assumed.
Results and discussion
Figure 3 An example of how fluorescent surface damage marks
on cotton bolls were selected and evaluated for analysis.
Strongly fluorescent regions, identified with a local maxima filter,
were analyzed with respect to area and intensity. One representative
region has been magnified, and the two auxiliary circular regions
indicated. Average intensity inside the smaller circle is denoted I1 , and
average intensity inside the larger circle is denoted I2 . In instances
where stink bugs or syringe needles made the puncture, the ratio of
I1 /I2 often falls below 1. On the other hand, when damage marks
were found on non-infested cotton bolls, the ratio of I1 /I2 was less
than 1 in only 24 % of cases.
The purpose of this study was to provide proof-ofprinciple that LED-induced fluorescence combined with
image analysis can provide an indicator whether a cotton
boll has been fed on by stink bugs or not. Here, the most
important component was the ability to detect stink bugrelated fluorescent marks on the outer carpal wall without
the need to manually open the boll.
The starting point for this study is an earlier publication [18], in which we reported on the observation of a
characteristic green fluorescence emission near stink bug
puncture sites. This fluorescence was prominently visible with the unaided eye under ultraviolet illumination on
the lint and the inner carpal wall. We also observed that
some fluorescence becomes visible on the outer carpal
wall, while at the same time the red chlorophyll emission
recedes.
Further examination of this phenomenon revealed that
the stink bug marks on the outer carpal wall are more
prominently visible under deep blue excitation than under
ultraviolet excitation: Contrast of verified puncture marks,
defined as the difference between green intensity and red
(chlorophyll) intensity normalized by red intensity, was
typically increased four-fold under deep blue illumination
when compared to UV illumination. For this reason, any
Mustafic et al. Journal of Biological Engineering 2013, 7:5
http://www.jbioleng.org/content/7/1/5
Page 5 of 9
further image acquisition was performed with LED excitation at 440nm. The fluorescence emission that accompanies typical damage on the inside of a cotton boll is shown
in Figure 4.
Any bolls that were punctured with a sterile needle
exhibited a similar fluorescent pattern on the outside
carpal wall and wart development on the inner carpal
wall, although the outside diameter of a 31 gauge needle
(0.26mm) is larger than a stink bug proboscis, estimated
at 0.17 mm [21]. Unlike in bolls punctured by stink bugs,
lint coloration and the later development of shriveled and
dried lint as well as mold was absent. We conclude that the
fluorescence is a by-product of the plant healing process,
but that stink bug feeding introduces additional damage
due to pathogen contamination. For the image analysis
process, however, this observation is of no relevance.
The image processing method was developed based on
the observation that (a) puncture marks cause a circular limited region where green fluorescence develops, and
(b) the center of the region, where the plant tissue was
removed by the puncture, actually exhibited very low fluorescence, leading to a donut-shaped intensity pattern.
Examples of the image patterns are shown in Figure 5,
both under normal fluorescence and after processing to
yield the ratiometric image.
One challenge for the image analysis method is the presence of fluorescent regions that are not related to stink
bug feeding (Figure 5E and F). In fact, out of 109 control cotton bolls that were not exposed to stink bugs,
41 showed fluorescent marks, and a total of 49 fluorescent marks met the inclusion criteria and were included
Figure 4 A fluorescent image of the cotton boll carpal wall (left)
and cotton boll lint (right) with arrows denoting the damage
caused by stink bugs. The carpal wall shows fluorescent interior
warts, which often have a noticeable dark center where the stink bug
pierced the wall. Stink bug damage to the lint is visible in two
different ways: lint browning (visible under regular white illumination)
and fluorescent damaged regions that correspond with the piercing
marks on the carpal wall.
Figure 5 Raw color images (left column) and corresponding
ratiometric images (right column) of cotton bolls with various
types of damage and stages of infestation. (A) and (B) show a
non-infested cotton boll without any exterior damage marks. (C) and
(D) show a cotton boll punctured by a sterilized syringe for negative
control studies. The black ring in (C) was drawn with a fiber tip
marker. (E) and (F) show an example of a control cotton boll that
exhibits fluorescent marks very similar to stink bug puncture marks.
Those marks would typically be identified as positives and count as
false positives. (G) and (H) are images of a cotton boll with exterior
puncture marks caused by stink bug feeding. The marks are visible as
larger fluorescent green regions and count as true positives. Note the
presence of smaller fluorescent dots not caused by stink bugs. The
scale bar (white line in lower left corner of image A) represents 5 mm.
Mustafic et al. Journal of Biological Engineering 2013, 7:5
http://www.jbioleng.org/content/7/1/5
Fluorescent area (mm2)
***
3
ns
***
2
1
P
N
IN
F
IN
F
0
N
in the analysis. Conversely, out of 136 cotton bolls that
were exposed to stink bugs, 38 did not show any fluorescent marks, and manual examination of the inner
carpal wall and lint confirmed that these bolls had not
been fed upon by the stink bugs (i.e., true negatives).
Of the remaining 98 cotton bolls that exhibited fluorescent marks, a total of 169 marks were included in
the analysis. Manual examination of cotton bolls not
included in this study led to the inclusion criteria, namely,
a minimum area, a minimum average intensity, and a
strictly convex shape. It is conceivable that small marks
are caused by non-piercing insects, such as spider mites
[22], and large, elongated marks are the consequence of
mechanical abrasion.
Clearly, any image analysis method must detect falsepositive fluorescent marks. The first step of the image
analysis chain was aimed at detecting all fluorescent
marks, irrespective of their origin. Local maximum detection combined with either gradient-based segmentation
(e.g., the hill-climbing algorithm [23] or active contours
(“snakes” [20]) or intensity-based segmentation, such as
region-growing, can be used in a straightforward manner to provide the initial set of candidates. Based on
the initial set of cotton bolls, we defined all fluorescent
marks that were found on the control group of cotton
bolls as false-positives, and all fluorescent marks that were
found on the exposed group of cotton bolls as potential true positives. The fluorescent marks associated with
needle punctures (total of 81 marks) were examined as a
separate group.
The second step of the image analysis process, which
was the focus of this study, was aimed at separating true
positives from false positives. Among several quantitative metrics, size had a strong ability to separate false and
true positives as shown in Figure 6. The median area of
the fluorescent region associated with stink bug feeding
was 0.35mm2 compared to a median area of 0.6mm2 for
the false positives. The difference was statistically significant (P < 0.0001). Needle puncture marks caused an
even smaller fluorescent region with a median area of
0.27mm2 . Size variation was less within the marks caused
by stink bug feeding and those caused by the needle puncture, whereas a large variability in size was observed in
the false-positive group. Median size was not statistically different between stink bug-related marks and needle
puncture marks.
A second quantitative metric was obtained by measuring the intensity drop-off towards the center of the puncture mark. False-positive marks tend to have an irregular
intensity distribution with the highest intensity near the
center. Conversely, plant tissue that was directly damaged by the puncture was almost black under visible-light
examination, and its corresponding fluorescent emission
lower. It can therefore be expected that an intensity ratio
Page 6 of 9
Figure 6 Boxplot of the area A of the segmented fluorescent
region for three categories of fluorescent spots: non-infested
(NINF), infested (INF), and needle punctures (NP). Median area is
significantly larger (P < 0.0001) for fluorescent areas that are neither
stink bug-related nor caused by needle puncture, but the area is not
significantly different between stink bug feeding marks and needle
punctures (Kruskal-Wallis test with Dunn’s multiple-comparison
post-test).
of the center intensity I1 , relative to the average intensity
of the entire fluorescent spot I2 , tends to be less than 1
for true-positives and greater than 1 for false-positives. In
some cases (such as the one displayed in Figure 5G and H),
the dark center region was not as pronounced as in most
cases, and we accepted an intensity ratio of I1 /I2 < 1.2 as
indication of a positive (i.e., stink bug-caused) fluorescent
mark.
A box plot of the intensity ratio can be seen in Figure 7.
Median ratiometric intensity of the false-positive fluorescent regions was 1.2, with a statistically significant
difference from 1 (Wilcoxon signed rank test, P < 0.0001).
Conversely, median ratiometric intensity of those fluorescent regions associated with stink bug damage was 0.96,
statistically significant lower than 1 (Wilcoxon signed rank
test, P < 0.0001). For comparison, needle punctures
exhibited a median ratiometric intensity of 0.86, lower
than 1 with P = 0.0002. Similar to size, the ratiometric
intensity was significantly different between false positives and true positives (P < 0.0001) and between false
positives and needle puncture marks (P < 0.0001).
It is interesting to note the relative similarity between
fluorescent marks caused by stink bug feeding and those
caused by needle puncture, which is consistent with our
earlier publication [18], where we hypothesized that the
fluorescence emission is the by-product of the plant healing process. Equally relevant is the observation that the
fluorescent marks do not fade over a period of seven days
Mustafic et al. Journal of Biological Engineering 2013, 7:5
http://www.jbioleng.org/content/7/1/5
Fluorescent area (mm2)
***
3
*
***
I1/I2
2
Page 7 of 9
1
3
2
1
0
0.0
N
P
IN
F
N
IN
F
0
Figure 7 Boxplot of the intensity ratio I1 /I2 for three categories of
fluorescent spots: non-infested (NINF), infested (INF), and needle
punctures (NP). Fluorescent marks that are not related to a puncture
generally have a higher intensity near the center. Correspondingly, the
ratio I1 /I2 is greater than 1 (statistically significant, Wilcoxon signed
rank test, P < 0.0001). Puncture marks tend to have a darker center,
and the ratio I1 /I2 is less than 1. This difference is also statistically
significant for stink bug feeding marks and needle punctures.
Furthermore, the ratio I1 /I2 is significantly higher for fluorescent
marks that are not related to a puncture than for those related to stink
bug feeding and those related to needle punctures (both P < 0.0001
by the Kruskal-Wallis test with Dunn’s multiple-comparison post-test).
(i.e., the maximum period over which we examined a cotton boll). This finding confirms those we reported earlier
[18].
Although both the size metric and the intensity ratio
metric are different between stink bug feeding marks and
unrelated fluorescent spots from a statistical perspective,
considerable overlap between the values of the true positive and false positive groups exists. We therefore examined if a two-dimensional separation can further improve
the separation of false and true positives. A scatter plot
of both groups in two dimensions is shown in Figure 8.
Overlap still exists, but it can be envisioned that a rectangle that separates the lower left corner from the rest of
the region can capture most of the true positives. Such
a rectangle could, for example, combine the conditions
I1 /I2 < 1.2 and A < 0.6. Puncture marks from infested
cotton bolls have a ratio I1 /I2 < 1 in 69% of the cases,
whereas for false-positive damage the ratio falls below one
in only 24% of the cases. Only four (8%) false-positives
lie within the rectangle arbitrarily defined above, and 147
true-positives (87%) lie within the same rectangle. Contingency tables for the criteria cited above are provided
in Table 1. Clearly, two-dimensional separation markedly
improves the specificity of the detection method.
In an attempt to determine the limits of separation
between true and false positives, a receiver operating characteristic (ROC) analysis was performed by increasing the
Infested
Non-infested
0.5
1.0
1.5
2.0
I1/I2
Figure 8 Scatterplot of all examined marks (except the needle
puncture controls) arranged in two dimensions by area A and
intensity ratio I1 /I2 . A multiple-threshold criterion, for example
I1 /I2 < 1.2 and A < 0.6, improves the separation of false-positive
marks from true-positive marks: 87% of the true-positive marks lie
inside the rectangle, whereas only 8% of the false-positives fall into
the same rectangular region.
size of the rectangle along a diagonal through the origin and the point (0.6, 1.2). As shown in Figure 9, choice
of a very small rectangle leads to high sensitivity, but
poor specificity due to the exclusion of many true positives. Too large a rectangle leads to poor sensitivity due
to the inclusion of the false positives. The ROC curve,
however is very different from the diagonal (i.e., assumption of random data). The ROC analysis does not take
into account that different separation boundaries (e.g., a
diagonal line connecting A = 1.1 and I1 /I2 = 1.5, or
an ellipse) can improve both sensitivity and specificity
even further. These considerations are not appropriate for
this stage of the research, however, because a larger-scale
study is necessary to identify data obtained in the field. It
is conceivable that clustering algorithms can provide the
boundary from a training data set, and points placed in
the two-dimensional coordinate system can be assigned to
their respective cluster.
For practical purposes, an 87% detection of true positives with only 8% false-positives is already acceptable
when we consider that the main goal of the application of
this method in the field is the approximate determination
of the level of stink bug infestation. The reason for this
conclusion lies in the study by Reay-Jones [24] who found
an average damage rate of 14.8% on a scale from 0% to
100% in a very large sample from commercial cotton fields
in Georgia and South Carolina. Moreover, Extensionrecommended treatment thresholds range from 15% to
30% internal boll damage, depending on week of bloom
[25]. In our previous study of the new fluorescent method
[18], accuracy for the conventional visual inspection was
found to be 75%, while the accuracy for the florescent
Mustafic et al. Journal of Biological Engineering 2013, 7:5
http://www.jbioleng.org/content/7/1/5
Page 8 of 9
Table 1 Contingency tables for separation of fluorescent marks by their area A, their intensity ratio I1 /I2 , and both
criteria simultaneously
A < 0.6
I1
I2
A ≥ 0.6
< 1.2
I1
I2
≥ 1.2
A < 0.6 and
I1
I2
< 1.2
A ≥ 0.6 or
I1
I2
≥ 1.2
Infested n = 169
70.2%
7.3%
73.9%
3.7%
67.4%
10.6%
Control n = 49
12.4%
10.1%
11.0%
11.5%
6.0%
16.5%
detection method was 92.9% with a 7.7% false positive rate
when a human observer analyzed the fluorescent images.
A concurrent development of a method to detect stink
bug presence and damage to cotton bolls is an ”electronic nose” sensor for specific volatiles emitted by stink
bugs or damaged plants [26]. Under laboratory conditions,
the internal boll injury (interior walls of bolls with warts
and lint damage) was predicted in 95% of all cases [26],
and the electronic nose was able to identify injured and
healthy cotton bolls with an accuracy between 80 and 90%
[27]. Notably, the detection accuracy of 84% reported in
this study is in a similar range as the one reported by
Degenhardt et al. [27].
In our previous work [18], we identified a fluorescent
wand and a fluorescent image scanner as possible devices
to detect stink bug-related fluorescence. Further analysis
of the fluorescent marks, as presented in this study, reveals
that the wand is likely impractical due to the presence of
a large number of false-positive fluorescent marks. The
second device that we proposed is a light-shielded enclosure in which fluorescence emission can be detected by
1
1.0
Conclusions
2
Sensitivity
3
4
0.5
5
6
0.0
0.0
0.5
pushing the cotton boll into the enclosure. With a suitably
engineered device, the cotton boll might even be left on
the plant. Irrespective of the final embodiment of a fielddeployable device, we can expect a significant increase of
cotton boll throughput and analysis speed. Based on the
study by Reay-Jones [24], the appropriate sample size for
arriving at a treatment decision is 15 samples (20 bolls
per sample) under the assumption of a boll injury rate of
14.3% and the accepted error rate of 30% of the mean.
Toews et al. [16] compared the time required for internal
examination of bolls and found that a single 20-boll sample required 7 minutes to collect and process. Since field
scouts are unable to invest the required time to reach this
level of precision, they tend to collect fewer bolls and consequently accept a much higher error rate. The fluorescent
detection promises much faster sampling as image acquisition and analyses could be performed instantaneously.
Examining more bolls in less time promises a more accurate estimate of boll damage. These data strongly support
that the fluorescent detection method is a major step
forward in making more accurate estimates of stink bug
damaged bolls for making pest management decisions.
1.0
1-Specificity
Figure 9 Receiver operating characteristic (ROC) for
different-size rectangles (cf. Figure 8) as a study how well a dual
threshold criterion can separate true and false positive
fluorescent marks. As specific corner points for the rectangle in
Figure 8 were chosen: 1: (1.5, 1.4), 2: (1.0, 1.2), 3: (0.8, 1.1), 4: (0.65, 1.05),
5: (0.5, 1.0), 6: (0.4, 0.98). At point 1, sensitivity is high but specificity is
low, and this changes towards the higher point number. At point 6
we reach lower values for sensitivity but higher values for specificity.
Striking a balance in this instance includes reducing the number of
false positives (high sensitivity) and false negatives (high specificity).
Dashed line represents a random assumption.
Among the many challenges facing agronomists in growing cotton, two stand prominently: detection of internal damage and determination of its underlying cause.
With respect to cotton, current methods include harvesting immature cotton bolls and opening them so
any damage can be visualized. Even if no damage is
observed, the cotton bolls are lost irretrievably due to
destructive sampling. Furthermore, small bolls are nearly
always shed by the plant as a result of stink bug feeding [28]. In that case, the damaged bolls would not be
included in the samples, because they would have been
dropped by the plant before reaching appropriate diameter for sampling. It is also desirable to detect boll damage very quickly so that the stink bug population can
be mitigated with insecticides before the bugs move to
new hosts.
To overcome these limitations, we present one step
towards a field applicable imaging method to estimate the
levels of stink bug infestation. The next step of this project
is the engineering design of an imaging scanner that can
be used in the field. We envision a hand-held imaging
apparatus, where the illumination source and an image
Mustafic et al. Journal of Biological Engineering 2013, 7:5
http://www.jbioleng.org/content/7/1/5
sensor array are placed in an enclosure. A cotton boll that
is placed in this enclosure needs to be shielded from environmental light. One desirable property of this enclosure
that allows the boll to stay on the plant, although sampling
by harvesting some bolls and inserting them in the imaging apparatus may turn out to be a suitable alternative.
An additional motivation for developing a device that can
analyze the boll on the plant is its use in research to examine fluorescent marks and stink bug damage over a longer
period of time.
Either way, with the image analysis methods described
herein, the theoretical foundations towards a field-usable
detection device for stink bug damage in cotton bolls
are laid.
Competing interests
The authors declare that they have no competing interests.
Authors’ contributions
MAH and MT made the original discovery of stinkbug-related fluorescence. MT
provided the controlled set of cotton bolls. MAH designed the imaging
apparatus and parts of the image analysis chain. AM and MAH developed the
intensity ratio metric. AM and EER acquired and processed all photographic
and microscopic images, and performed the statistical analysis. All authors
contributed to the data analysis and to the manuscript. All authors read and
approved the final manuscript.
Author details
1 College of Engineering, University of Georgia, Athens, GA 30602-4435, USA.
2 Department of Entomology, University of Georgia, Tifton, GA 31793-0748,
USA.
Received: 14 June 2012 Accepted: 11 January 2013
Published: 20 February 2013
References
1. Greene JK, Bundy CS, Roberts PM, Leonard BR: Identification and
management of common boll feeding bugs in cotton. Clemson
Extension, Report, EB158, Clemson University, Blackville, SC, USA 2006.
2. Panizzi AR: Wild hosts of pentatomids: ecological significance and
role in their pest status on crops. Annu Rev Entomol 1997, 42:99–122.
3. Mitchell WC, Warner RM, Fukunaga ET: Southern green stink bug,
Nezara viridula (L.), injury to macadamia nut. Proc Hawaiian Entomol
Soc 1965, 19:103–109.
4. McPherson J, McPherson R: Stink Bugs of Economic Importance in America
North of Mexico: CRC; 2000.
5. Bommireddy PL, Leonard BR, Temple JH: Influence of Nezara viridula
feeding on cotton yield, fiber quality, and seed germination. J Econ
Entomol 2007, 100(5):1560–1568.
6. Medrano E, Esquivel J, Bell A: Transmission of cotton seed and boll
rotting bacteria by the southern green stink bug (Nezara viridula L.).
J Appl Microbiol 2007, 103(2):436–444.
7. Medrano EG, Esquivel JF, Nichols RL, Bell AA: Temporal analysis of
cotton boll symptoms resulting from southern green stink bug
feeding and transmission of a bacterial pathogen. J Econ Entomol
2009, 102:36–42.
8. Toscano N, Stern V: Cotton yield and quality loss caused by various
levels of stink bug infestations. J Econ Entomol 1976, 69:53–56.
9. Barbour K, Bradley J, Julius R, Bacheler J: Reduction in yield and quality
of cotton damaged by green stink bug (Hemiptera: Pentatomidae).
J Econ Entomol 1990, 83(3):842–845.
10. Greene J, Turnipseed S, Sullivan M, Herzog G: Boll damage by southern
green stink bug (Hemiptera: Pentatomidae) and tarnished plant
bug (Hemiptera: Miridae) caged on transgenic Bacillus thuringiensis
cotton. J Econ Entomol 1999, 92(4):941–944.
Page 9 of 9
11. Willrich MM, Leonard BR, Temple J: Injury to preflowering and
flowering cotton by brown stink bug and southern green stink bug.
J Econ Entomol 2004, 97(3):924–933.
12. Onstad D: Calculation of economic-injury levels and economic
thresholds for pest management. J Econ Entomol 1987, 80(2):297–303.
13. Chi H: Timing of control based on the stage structure of pest
populations: a simulation approach. J Econ Entomol 1990,
83(4):1143–1150.
14. Anonymous: Cotton: World Markets and Trade. USDA Circular Series
FOP 11 2012. Available online: http://usda01.library.cornell.edu/usda/fas/
cotton-market//2010s/201%2/cotton-market-11-09-2012.pdf.
15. Anonymous: Cotton Council Internaltional 2009 Buyer’s Guide.
Available online: http://www.cottonusa.org/directories/buyersguide.cfm?
itemnumber=1545%&sn.itemnumber=1080&tn.itemnumber=1105.
16. Toews M, Greene J, Reay-Jones F, Reeves R, Boyd S, Huffman M, Richter D,
Robertson B: A comparison of sampling techniques for stink bugs in
cotton. In Proceedings of the beltwide cotton conferences; 2008:1193–1203.
17. Reay-Jones F, Toews M, Greene J, Reeves R: Spatial dynamics of stink
bugs (Hemiptera: Pentatomidae) and associated boll injury in
southeastern cotton fields. Environ Entomol 2010, 39(3):956–969.
18. Xia J, Mustafic A, Toews M, Haidekker M: Stink bug feeding induces
fluorescence in developing cotton bolls. J Biol Eng 2011, 5:11.
19. Harris V, Todd J: Comparative fecundity egg fertility and hatch aong
wild-type and three laboratory-reared generations of the southern
green stink bug, Nezara viridula (L.) (Hemiptera: Pentatomidae). J
Georgia Entomol Soc 1980, 15(3):241–245.
20. Haidekker MA: Advanced Biomedical Image Analysis: Wiley & Sons; 2011.
21. Leske T, Short B, Wright S, Brown M: Diagnosis and variation in
appearance of brown stink bug (Hemiptera: Pentatomidae) injury
on apple. J Entomol Sci 2009, 44:314–322.
22. Smith C, Cothren J: Cotton – Origin, History, Technology, and Production:
Wiley Series in Crop Science; 1999.
23. Bankman IN, Nizialek T, Simon I, Gatewood OB, Weinberg IN, Brody WR:
Segmentation algorithms for detecting microcalcifications in
mammograms. IEEE Trans Inf Technol Biomed 1997, 1(2):141–149.
24. Reay-Jones F, Toews M, Greene J, Reeves R: Development of sampling
plans for cotton bolls injured by stink bugs (Hemiptera:
Pentatomidae). J Econ Entomol 2010, 103(2):525–532.
25. Bacheler J, Herbert DA, Greene J, Roberts P, Toews M, Blinka E: Scouting
for stink bug damage in southeast cotton: description and use of a
pocket scouting decision aid. North Carolina State University Cooperative
Extension 2010. Available online http://ipm.ncsu.edu/cotton/
insectcorner/pdf/AG 730 WPrint-NC.pdf.
26. Henderson WG, Khalilian A, Han YJ, Greene JK, Degenhardt DC:
Detecting stink bugs/damage in cotton utilizing a portable
electronic nose. Comput Electron Agric 2010, 70:157–162.
27. Degenhardt DC, Greene JK, Khalilian A: Temporal dynamics and
electronic nose detection of stink bug-induced volatile emissions
from cotton bolls. Psyche: A J Entomol 2012.
28. Huang TI, Toews M: Feeding preference and movement of Nezara
viridula (L.) and Euschistus servus (Say) (Hemiptera: Pentatomidae)
on individual cotton plants. J Econ Entomol 2012, 105:847–853.
doi:10.1186/1754-1611-7-5
Cite this article as: Mustafic et al.: LED-Induced fluorescence and image
analysis to detect stink bug damage in cotton bolls. Journal of Biological
Engineering 2013 7:5.